diff --git a/ggml.c b/ggml.c index c0e7ec05c..c24992260 100644 --- a/ggml.c +++ b/ggml.c @@ -3494,7 +3494,7 @@ static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = { }; static_assert(GGML_TYPE_COUNT == 13, "GGML_IS_QUANTIZED is outdated"); -static const char * GGML_OP_LABEL[GGML_OP_COUNT] = { +static const char * GGML_OP_NAME[GGML_OP_COUNT] = { "NONE", "DUP", @@ -3749,6 +3749,9 @@ const char * ggml_type_name(enum ggml_type type) { return GGML_TYPE_NAME[type]; } +const char * ggml_op_name(enum ggml_op op) { + return GGML_OP_NAME[op]; +} size_t ggml_element_size(const struct ggml_tensor * tensor) { return GGML_TYPE_SIZE[tensor->type]; @@ -4017,6 +4020,10 @@ size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch) return result; } +void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc) { + ctx->no_alloc = no_alloc; +} + // IMPORTANT: // when creating "opt" tensors, always save and load the scratch buffer // this is an error prone process, but it is necessary to support inplace @@ -4061,7 +4068,7 @@ struct ggml_tensor * ggml_new_tensor_impl( struct ggml_object * const obj_new = (struct ggml_object *)(mem_buffer + cur_end); if (ctx->scratch.data == NULL || data != NULL) { - size_needed += sizeof(struct ggml_tensor); + size_needed += GGML_TENSOR_SIZE; if (cur_end + size_needed + GGML_OBJECT_SIZE > ctx->mem_size) { GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", @@ -4077,14 +4084,15 @@ struct ggml_tensor * ggml_new_tensor_impl( }; } else { if (ctx->scratch.offs + size_needed > ctx->scratch.size) { - GGML_PRINT("%s: not enough space in the scratch memory\n", __func__); + GGML_PRINT("%s: not enough space in the scratch memory pool (needed %zu, available %zu)\n", + __func__, ctx->scratch.offs + size_needed, ctx->scratch.size); assert(false); return NULL; } - if (cur_end + sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE > ctx->mem_size) { + if (cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE > ctx->mem_size) { GGML_PRINT("%s: not enough space in the context's memory pool (needed %zu, available %zu)\n", - __func__, cur_end + sizeof(struct ggml_tensor) + GGML_OBJECT_SIZE, ctx->mem_size); + __func__, cur_end + GGML_TENSOR_SIZE + GGML_OBJECT_SIZE, ctx->mem_size); assert(false); return NULL; } @@ -4093,7 +4101,7 @@ struct ggml_tensor * ggml_new_tensor_impl( *obj_new = (struct ggml_object) { .offs = cur_end + GGML_OBJECT_SIZE, - .size = sizeof(struct ggml_tensor), + .size = GGML_TENSOR_SIZE, .next = NULL, }; @@ -13792,11 +13800,19 @@ static void ggml_visit_parents(struct ggml_cgraph * cgraph, struct ggml_tensor * // reached a leaf node, not part of the gradient graph (e.g. a constant) GGML_ASSERT(cgraph->n_leafs < GGML_MAX_NODES); + if (strlen(node->name) == 0) { + snprintf(node->name, sizeof(node->name), "leaf_%d", cgraph->n_leafs); + } + cgraph->leafs[cgraph->n_leafs] = node; cgraph->n_leafs++; } else { GGML_ASSERT(cgraph->n_nodes < GGML_MAX_NODES); + if (strlen(node->name) == 0) { + snprintf(node->name, sizeof(node->name), "node_%d", cgraph->n_nodes); + } + cgraph->nodes[cgraph->n_nodes] = node; cgraph->grads[cgraph->n_nodes] = node->grad; cgraph->n_nodes++; @@ -14510,6 +14526,18 @@ void ggml_graph_reset(struct ggml_cgraph * cgraph) { } } +struct ggml_tensor * ggml_get_tensor_by_name(struct ggml_cgraph * cgraph, const char * name) { + for (int i = 0; i < cgraph->n_nodes; i++) { + struct ggml_tensor * node = cgraph->nodes[i]; + + if (strcmp(node->name, name) == 0) { + return node; + } + } + + return NULL; +} + void ggml_graph_print(const struct ggml_cgraph * cgraph) { int64_t perf_total_per_op_us[GGML_OP_COUNT] = {0}; @@ -14527,7 +14555,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 ", %5" PRId64 "] %16s %s (%3d) cpu = %7.3f / %7.3f ms, wall = %7.3f / %7.3f ms\n", i, node->ne[0], node->ne[1], node->ne[2], - GGML_OP_LABEL[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs, + GGML_OP_NAME[node->op], node->is_param ? "x" : node->grad ? "g" : " ", node->perf_runs, (double) node->perf_cycles / (double) ggml_cycles_per_ms(), (double) node->perf_cycles / (double) ggml_cycles_per_ms() / (double) node->perf_runs, (double) node->perf_time_us / 1000.0, @@ -14541,7 +14569,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { GGML_PRINT(" - %3d: [ %5" PRId64 ", %5" PRId64 "] %8s\n", i, node->ne[0], node->ne[1], - GGML_OP_LABEL[node->op]); + GGML_OP_NAME[node->op]); } for (int i = 0; i < GGML_OP_COUNT; i++) { @@ -14549,7 +14577,7 @@ void ggml_graph_print(const struct ggml_cgraph * cgraph) { continue; } - GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_LABEL[i], (double) perf_total_per_op_us[i] / 1000.0); + GGML_PRINT("perf_total_per_op_us[%16s] = %7.3f ms\n", GGML_OP_NAME[i], (double) perf_total_per_op_us[i] / 1000.0); } GGML_PRINT("========================================\n"); diff --git a/ggml.h b/ggml.h index c22d93836..0c90f5064 100644 --- a/ggml.h +++ b/ggml.h @@ -198,6 +198,7 @@ #define GGML_MAX_PARAMS 256 #define GGML_MAX_CONTEXTS 64 #define GGML_MAX_OPT 4 +#define GGML_MAX_NAME 32 #define GGML_DEFAULT_N_THREADS 4 #define GGML_ASSERT(x) \ @@ -372,11 +373,16 @@ extern "C" { void * data; - char name[32]; + char name[GGML_MAX_NAME]; char padding[16]; }; + static const size_t GGML_TENSOR_SIZE = sizeof(struct ggml_tensor); + + // use this to compute the memory overhead of a tensor + static const size_t GGML_TENSOR_OVERHEAD = (GGML_OBJECT_SIZE + GGML_TENSOR_SIZE + 16); + // computation graph struct ggml_cgraph { int n_nodes; @@ -429,6 +435,7 @@ extern "C" { GGML_API float ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float GGML_API const char * ggml_type_name(enum ggml_type type); + GGML_API const char * ggml_op_name (enum ggml_op op); GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor); @@ -445,6 +452,7 @@ extern "C" { GGML_API size_t ggml_used_mem(const struct ggml_context * ctx); GGML_API size_t ggml_set_scratch(struct ggml_context * ctx, struct ggml_scratch scratch); + GGML_API void ggml_set_no_alloc(struct ggml_context * ctx, bool no_alloc); GGML_API struct ggml_tensor * ggml_new_tensor( struct ggml_context * ctx, @@ -970,6 +978,8 @@ extern "C" { GGML_API void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph); GGML_API void ggml_graph_reset (struct ggml_cgraph * cgraph); + GGML_API struct ggml_tensor * ggml_get_tensor_by_name(struct ggml_cgraph * cgraph, const char * name); + // print info and performance information for the graph GGML_API void ggml_graph_print(const struct ggml_cgraph * cgraph);